DistVS: Large-scale Vector Search with Compute-Memory Disaggregation
summary
DistVS maps low precision PQ into the compute tier, high precision PQ plus the graph into a memory tier, and full vectors on SSDs, and executes an iterative low→high→full search called PRESS that prunes candidates before expensive IO. System optimizations include asynchronous execution and RDMA IO batching and decoupled re‑ranking. Evaluation on 100M–1B vectors (dims 96–768) shows usually over 40% higher throughput versus prior disk‑based and distributed systems and up to 114% in some cases.
AI-written summary · about
chapters (8)
- 0:00DistVS and vector search
- 3:28Two-tier vector search challenges
- 8:40Compute-memory disaggregation
- 10:25Three-precision algorithm
- 11:52DistVS distributed design
- 14:02System optimizations
- 15:26Evaluation and results
- 16:50Conclusion
related talks
12mVectorCDC: Accelerating Data Deduplication with Vector Instructions
37mMilvus: Scaling Vector Data Solutions with Gen AI
01h 02mSqueezing the Hardware to Make Performance Juice
28mSearch and Sushi; Freshness Counts
09mPD Disaggregation VLLM Deployment on Alternative AI Accelerators Using Llm-d
01h 04mDesign for Performance
01h 01mFaster, Easier, Simpler Vectors
07mTest-Drive Performance